Abstract
Using a large deviations approach we calculate the probability distribution of the mutual information of MIMO channels in the limit of large antenna numbers. In contrast to previous methods that only focused at the distribution close to its mean (thus obtaining an asymptotically Gaussian distribution), we calculate the full distribution, including its tails which strongly deviate from the Gaussian behavior near the mean. The resulting distribution interpolates seamlessly between the Gaussian approximation for rates $R$ close to the ergodic value of the mutual information and the approach of Zheng and Tse for large signal to noise ratios $ ho$. This calculation provides us with a tool to obtain outage probabilities analytically at any point in the $(R, \rho, N)$ parameter space, as long as the number of antennas $N$ is not too small. In addition, this method also yields the probability distribution of eigenvalues constrained in the subspace where the mutual information per antenna is fixed to $R$ for a given $ ho$. Quite remarkably, this eigenvalue density is of the form of the Marcenko-Pastur distribution with square-root singularities, and it depends on the values of $R$ and $ ho$.
The Challenge of Outage Capacity
In Multiple-Input Multiple-Output (MIMO) wireless systems, a critical engineering requirement is understanding "outage capacity"—the probability that a fading channel cannot support a target data rate. Calculating this requires the full probability distribution of the channel's mutual information.
Historically, researchers have relied on two main approximations. The Gaussian approximation works well near the average (ergodic) capacity but fails significantly at the "tails" of the distribution. Since reliable communication systems operate at low outage probabilities (e.g., 1% or 0.1%), these tails are precisely where accuracy is most needed. Conversely, the large-SNR approximation captures the tails but is only valid when the signal-to-noise ratio is extremely high. This research bridges the gap between these two regimes, providing a single analytical framework that predicts outage probabilities across all rates and SNRs.
Methodology: The Coulomb Gas Analogy
The authors employ a Large Deviations approach rooted in statistical physics, specifically the Coulomb gas analogy originally proposed by Freeman Dyson. Instead of direct matrix algebra, the eigenvalues of the MIMO channel matrix (the Wishart matrix $\mathbf{H}^\dagger\mathbf{H}$) are modeled as a gas of positively charged particles on a 1D line.
In this model, particles repel each other logarithmically and are held in place by an external potential field representing the matrix's properties. Determining the probability of a specific mutual information rate involves finding the minimum energy configuration of this charged gas subject to a rate constraint. By solving the resulting integral equations, the authors identify the most likely distribution of eigenvalues even at the extreme tails of the capacity distribution.
Key Results and Insights
Bridging the Asymptotic Gap
The Coulomb gas method seamlessly interpolates between the Gaussian approximation (valid for small deviations from the mean) and large-SNR asymptotic limits, such as the Zheng-Tse diversity-multiplexing tradeoff. This proves that these previously disconnected theories are special cases of a broader framework.
Generalized Marčenko-Pastur Distribution
While the standard Marčenko-Pastur law describes the average eigenvalue distribution of large random matrices, this work derives a generalized version. This distribution shows how channel eigenvalues rearrange themselves when forced to support a rate far from the average.
As shown in the figure above, the analytical model (dashed line) accurately tracks Monte Carlo simulations even for small antenna arrays (e.g., 2x4 or 3x6). This "unreasonable effectiveness" at small scales allows for accurate outage formulas in the $10^{-4}$ to $10^{-1}$ regime, where Gaussian approximations typically fail.
Phase Transitions and Diversity
The authors identified a "phase transition" in the eigenvalue distribution depending on the target rate and SNR. Furthermore, the model successfully recovers the "full diversity" scaling behavior at low rates, proving that outage probability scales as $\rho^{-MN}$ (where $\rho$ is SNR, and $M, N$ are the number of antennas).
Limitations and Future Work
The current model assumes idealized independent and identically distributed (i.i.d.) Gaussian noise and input, neglecting spatial correlation between antennas found in real-world hardware. Future work aims to extend this Coulomb gas methodology to correlated channels, which involves significantly more complex mathematical tools. Additionally, because the method yields the exact eigenvalue density for a required outage probability, it could be used to optimize transmission rates via algorithms like waterfilling when channel state information is available at the transmitter.
Summary updated